Transformer-based unsupervised contrastive learning for histopathological image classification.

Area of Science:

• Digital pathology
• Machine learning
• Computer vision

Background:

• Deep learning in medical imaging requires large annotated datasets, which are difficult to obtain for histopathology due to image complexity.
• Self-supervised learning (SSL) offers a solution by leveraging unlabeled data to create informative representations.

Purpose of the Study:

• To introduce a novel SSL strategy, semantically-relevant contrastive learning (SRCL), for histopathological images.
• To develop a hybrid deep learning model (CTransPath) for feature extraction.
• To evaluate the effectiveness of SRCL-pretrained CTransPath on various downstream tasks.

Main Methods:

• Proposed SRCL strategy to mine more positive pairs by comparing instance relevance, increasing representation diversity.
• Developed CTransPath, a hybrid CNN-Swin Transformer model, pretrained on unlabeled histopathology images.
• Evaluated performance on five downstream tasks across nine public datasets.

Main Results:

• SRCL-pretrained CTransPath achieved state-of-the-art performance on all tested downstream tasks.
• The learned visual representations demonstrated superior robustness and transferability compared to other SSL methods and ImageNet pretraining.
• The method proved effective across patch retrieval, classification, whole-slide image classification, mitosis detection, and gland segmentation.

Conclusions:

• SRCL is a highly effective SSL strategy for histopathology image analysis.
• The CTransPath model, pretrained with SRCL, provides powerful and versatile feature representations.
• This approach significantly advances deep learning applications in digital pathology, especially with limited annotations.